AIF-C01 Domain 1: Fundamentals of AI and ML
Part 2 of 6 in the AIF-C01 exam prep series for developers. Previous: The exam and what changed in v1.1.
Domain 1 is the vocabulary layer of the exam. Most of it is recognition: which term is which, which learning type fits a scenario, which AWS service matches a use case. It's also where developers lose easy points by skimming, because the definitions feel obvious until two of them appear side by side as answer options.
Domain at a glance#
Domain 1 is 20% of scored content, about 10 of the 50 scored questions.
| Task | What it covers |
|---|---|
| 1.1 | Explain basic AI concepts and terminology |
| 1.2 | Identify practical use cases for AI |
| 1.3 | Describe the AI/ML development lifecycle |
Objective checklist#
| # | Objective | Section |
|---|---|---|
| 1.1.1 | Define basic AI terms, including GenAI and agentic AI | Key terms |
| 1.1.2 | Differences between AI, ML, deep learning, GenAI, and agentic AI | How the terms nest |
| 1.1.3 | Inference types: batch, real-time, asynchronous, serverless | Inference types |
| 1.1.4 | Data types: labeled/unlabeled, tabular, time-series, image, text, structured/unstructured | Data types |
| 1.1.5 | Learning types: supervised, unsupervised, reinforcement learning | Learning types |
| 1.2.1 | Where AI/ML adds value | When AI helps and when it doesn't |
| 1.2.2 | When AI/ML is not appropriate | When AI helps and when it doesn't |
| 1.2.3 | Regression, classification, clustering | Choosing an ML technique |
| 1.2.4 | Real-world AI applications, including knowledge bases and agentic AI | Real-world applications |
| 1.2.5 | Capabilities of AWS managed AI/ML services | AWS managed AI services |
| 1.2.6 | Traditional ML vs. foundation models for a use case | Traditional ML vs. foundation models |
| 1.3.1 | Components of an AI/ML pipeline | The ML pipeline |
| 1.3.2 | Sources of foundation models | Where models come from |
| 1.3.3 | Using a model in production: managed API vs. self-hosted | Where models come from |
| 1.3.4 | AWS services for each pipeline stage | AWS services by pipeline stage |
| 1.3.5 | MLOps fundamentals | MLOps |
| 1.3.6 | Model performance metrics and business metrics | Metrics |
Basic concepts and terminology#
Key terms#
| Term | Definition |
|---|---|
| Artificial intelligence (AI) | Systems that perform tasks normally requiring human intelligence: perceiving, reasoning, deciding |
| Machine learning (ML) | AI that learns patterns from data instead of following hand-written rules |
| Deep learning | ML using neural networks with many layers |
| Neural network | Layers of connected nodes whose weights are adjusted during training |
| Computer vision | AI that interprets images and video |
| Natural language processing (NLP) | AI that understands and generates human language |
| Algorithm | The learning procedure (for example, linear regression or a decision tree) |
| Model | The output of training an algorithm on data; the thing you run predictions with |
| Training | Adjusting a model's parameters using data |
| Inference | Using a trained model to make predictions on new data |
| Bias | Systematic error. In fairness terms, unfair outcomes for certain groups; in model-fit terms, a model too simple to capture the pattern |
| Fairness | Outcomes that don't discriminate against individuals or groups |
| Fit | How well a model captures the pattern in data: underfit, good fit, overfit |
| Large language model (LLM) | A very large transformer model trained on text to understand and generate language |
| Generative AI (GenAI) | Models that create new content: text, images, audio, video, code |
| Agentic AI | Systems that use a model to plan, call tools, and take actions toward a goal across multiple steps |
How the terms nest#
AI, ML, deep learning, and generative AI nest inside each other: each is a subset of the one before. Agentic AI is different. It's not a smaller circle inside GenAI; it's a system built on top of a generative model. It adds tools, memory, and an orchestration loop around the model.
Figure: AI ⊃ ML ⊃ deep learning ⊃ GenAI. Agentic AI wraps a generative model with tools, memory, and a loop.
Inference types#
Objective 1.1.3 now names four inference types. The examples below use SageMaker AI, where all four exist as deployment options.
| Type | How it works | Choose it when |
|---|---|---|
| Real-time | A persistent endpoint returns a prediction in milliseconds to seconds | Users are waiting for the answer and traffic is steady |
| Serverless | Endpoint scales automatically, down to zero when idle; the first request after idle may hit a cold start | Traffic is intermittent or unpredictable and occasional extra latency is acceptable |
| Asynchronous | Requests go into a queue and results are written to S3 when ready; handles large payloads (up to about 1 GB) and long processing (up to about an hour) | One request is big or slow, and nobody needs the answer instantly |
| Batch | Runs over an entire dataset offline; no endpoint stays up | You need predictions for a whole dataset on a schedule |
Figure: A quick decision path for the four inference types.
Data types#
| Distinction | Meaning | Example |
|---|---|---|
| Labeled | Each record has the correct answer attached | Emails tagged "spam" or "not spam" |
| Unlabeled | Raw data with no answers | A folder of customer reviews |
| Structured | Organized in rows and columns | A customer table in a database |
| Tabular | Structured data in a table | Loan applications with income, age, amount |
| Time-series | Values ordered by time | Hourly sales, sensor readings, stock prices |
| Unstructured | No predefined schema | Text documents, images, audio, video |
| Text | Unstructured language data | Support tickets, contracts |
| Image | Pixel data | Product photos, X-rays |
Labeling is expensive, which is why techniques that need little or no labeled data (unsupervised and self-supervised learning) matter so much for foundation models.
Learning types#
| Type | Learns from | Typical tasks | Example |
|---|---|---|---|
| Supervised | Labeled data | Regression, classification | Predict house prices; flag fraudulent transactions |
| Unsupervised | Unlabeled data | Clustering, anomaly detection, association, dimensionality reduction | Segment customers by behavior |
| Semi-supervised | A little labeled data plus a lot of unlabeled data | Classification when labels are scarce | Label 1% of documents, then learn from the rest |
| Self-supervised | Unlabeled data that generates its own labels | Pre-training foundation models | Predict the next word in a sentence |
| Reinforcement learning (RL) | Rewards and penalties from an environment | Sequential decision-making | Robotics, game playing, route optimization |
Reinforcement learning has its own vocabulary: an agent takes actions in an environment, observes the new state, and receives a reward. Over many episodes it learns a policy that maximizes cumulative reward.
Reinforcement learning from human feedback (RLHF) applies this to language models: humans rank model outputs, a reward model learns those preferences, and the language model is tuned toward them. It comes up again in Post 4 as part of fine-tuning.
Practical use cases#
When AI helps and when it doesn't#
AI and ML add value when you need to:
- Assist human decisions: triage, recommendations, risk scores.
- Scale beyond human capacity: reviewing millions of transactions or documents.
- Automate repetitive judgment: routing tickets, extracting fields from forms.
- Find patterns humans miss: anomalies, subtle correlations.
AI and ML are not appropriate when:
- A specific, deterministic outcome is required. Calculating tax, applying a fixed discount table, or enforcing a business rule should be code, not a prediction.
- The cost outweighs the benefit. Data collection, training, hosting, and monitoring can exceed the value of the improvement.
- There isn't enough quality data to learn the pattern.
- Every decision must be fully explainable, and no interpretable approach meets the bar.
- Errors are unacceptable and there's no room for human review.
Choosing an ML technique#
| Technique | Output | Learning type | Example |
|---|---|---|---|
| Regression | A continuous number | Supervised | Forecast next month's revenue; predict delivery time |
| Classification | A category (binary or multi-class) | Supervised | Spam or not spam; which product category |
| Clustering | Groups of similar items, with no predefined labels | Unsupervised | Customer segmentation |
Real-world applications#
| Application | What it does | AWS example |
|---|---|---|
| Computer vision | Detects objects, faces, text, and unsafe content in images and video | Amazon Rekognition |
| NLP | Extracts meaning, sentiment, and entities from text | Amazon Comprehend |
| Speech recognition | Converts speech to text | Amazon Transcribe |
| Recommendation systems | Suggests items based on behavior | Amazon Personalize |
| Fraud detection | Flags anomalous transactions | Custom models on SageMaker AI |
| Forecasting | Predicts future values from time-series data | Custom models on SageMaker AI |
| Knowledge bases | Answer questions grounded in your documents | Amazon Bedrock Knowledge Bases |
| Agentic AI | Plans and executes multi-step tasks with tools | Amazon Bedrock AgentCore |
AWS managed AI services#
These are pre-trained, API-driven services. You don't train anything; you call an API. Expect at least one matching question built from this table.
| Service | Input → output | Use it for |
|---|---|---|
| Amazon Comprehend | Text → insights | Sentiment, entities, key phrases, language detection, PII detection, topic modeling, custom classification |
| Amazon Transcribe | Speech → text | Call transcripts, subtitles, meeting notes; supports custom vocabularies and PII redaction |
| Amazon Translate | Text → text in another language | Localizing content and support conversations; supports custom terminology |
| Amazon Polly | Text → speech | Voice output for apps, reading content aloud; SSML controls pronunciation and pacing |
| Amazon Lex | Voice or text conversation → intent and slots | Chatbots and voice bots; fulfills requests through AWS Lambda |
| Amazon Rekognition | Images and video → labels | Object and face detection, content moderation, text in images, custom labels |
| Amazon Textract | Scanned documents → structured data | Extracting text, forms (key-value pairs), and tables from PDFs and images |
| Amazon Personalize | User behavior → recommendations | Product recommendations, personalized rankings |
| Amazon SageMaker AI | Your data → your own model | Building, training, and deploying custom ML models |
Traditional ML vs. foundation models#
Objective 1.2.6 is new, and the exam frames it around constraints rather than capabilities.
| Choose traditional ML when… | Choose a foundation model when… |
|---|---|
| A regulator requires you to explain each decision | The task involves understanding or generating language, images, or code |
| The data is tabular with a clear numeric or categorical target | You have few or no labeled examples |
| You need deterministic, repeatable behavior | One model must handle many varied tasks |
| Latency or cost per prediction must be very low | Time to market matters more than per-call cost |
| You have good labeled training data | The inputs are open-ended and hard to enumerate |
The AI/ML development lifecycle#
The ML pipeline#
Figure: The ML pipeline, a likely ordering question. Monitoring feeds back into data collection and retraining.
| Stage | What happens |
|---|---|
| Business goal and problem framing | Define success, then decide whether ML fits and which technique applies |
| Data collection | Gather data from sources; label it if the task is supervised |
| Pre-processing and exploratory data analysis (EDA) | Clean, fix missing values, remove duplicates; visualize distributions and correlations |
| Feature engineering | Select, transform, and create the input variables the model learns from |
| Training | Fit the algorithm to the training data |
| Hyperparameter tuning | Adjust settings you choose before training (learning rate, batch size, epochs, tree depth) |
| Evaluation | Measure performance on held-out data |
| Deployment | Serve the model with one of the inference types above |
| Monitoring | Watch for data drift and quality degradation; trigger retraining |
Datasets are usually split into training (fit the model), validation (tune hyperparameters), and test (final, unbiased evaluation) sets.
Where models come from and how they run#
Sources of foundation models (1.3.2):
- Pre-trained open-source models you download and host, for example from SageMaker JumpStart or Hugging Face.
- Proprietary models accessed through an API, for example through Amazon Bedrock.
- Custom models you train yourself. This is rarely justified for foundation models because of the data and compute required.
Using a model in production (1.3.3):
| Managed API service | Self-hosted | |
|---|---|---|
| AWS example | Amazon Bedrock | SageMaker AI endpoints, or EC2/EKS |
| You manage | Prompts, data, configuration | Instances, scaling, patching, the endpoint itself |
| Pricing | Per token or per request | Per instance-hour while the endpoint runs |
| Best for | Speed, simplicity, no infrastructure | Full control, custom or open-source models, specific hardware |
AWS services by pipeline stage#
| Stage | Services |
|---|---|
| Data preparation | SageMaker Data Wrangler, AWS Glue, AWS Glue DataBrew |
| Feature storage | SageMaker Feature Store |
| Training and tuning | SageMaker AI training jobs and automatic model tuning; SageMaker JumpStart for pre-trained models |
| Evaluation and bias checks | SageMaker Clarify; Amazon Bedrock Model Evaluation |
| Deployment | SageMaker AI endpoints; Amazon Bedrock for foundation models |
| Monitoring | SageMaker Model Monitor; Amazon CloudWatch |
| Orchestration and versioning | SageMaker Pipelines; SageMaker Model Registry |
| GenAI applications | Amazon Bedrock |
| Business users working with data and AI | Amazon Quick |
| Developers building with AI assistance | Kiro |
MLOps#
MLOps applies DevOps discipline to ML. The guide names these concepts:
| Concept | What it means in practice |
|---|---|
| Experimentation | Track runs, parameters, and results so you can compare and reproduce them |
| Repeatable processes | Automated pipelines instead of notebooks run by hand |
| Scalable systems | Infrastructure that handles growing data and traffic |
| Managing technical debt | Avoiding fragile glue code, undocumented features, and hidden dependencies |
| Production readiness | Testing, versioning, security, and rollback before go-live |
| Model monitoring | Detecting data drift, concept drift, and quality degradation |
| Model re-training | Retraining automatically or on a schedule when monitoring flags a problem |
Metrics#
Classification metrics start from the confusion matrix:
| Predicted positive | Predicted negative | |
|---|---|---|
| Actually positive | True positive (TP) | False negative (FN) |
| Actually negative | False positive (FP) | True negative (TN) |
| Metric | Question it answers | Optimize it when |
|---|---|---|
| Accuracy | What fraction of all predictions were right? | Classes are balanced |
| Precision | Of the items flagged positive, how many really were? | False positives are costly (a spam filter deleting real email) |
| Recall | Of the real positives, how many did we catch? | False negatives are costly (missing a disease or a fraud case) |
| F1 score | Balance of precision and recall | Classes are imbalanced and you need one number |
For regression, common metrics are mean absolute error (MAE), root mean squared error (RMSE), and R², which measures how much of the variance the model explains.
Business metrics answer a different question: is the model worth it? The guide lists cost per user, development costs, customer feedback, and return on investment (ROI).
Service cheat sheet#
| Service | One line |
|---|---|
| Amazon SageMaker AI | Build, train, tune, deploy, and monitor your own ML models |
| SageMaker JumpStart | Hub of pre-trained models you deploy or fine-tune on SageMaker |
| SageMaker Data Wrangler | Visual data preparation for ML |
| SageMaker Feature Store | Central store of ML features for training and inference |
| SageMaker Clarify | Bias detection and feature-attribution explainability |
| SageMaker Model Monitor | Detects data and quality drift in deployed models |
| Amazon Bedrock | Serverless API access to foundation models |
| Amazon Comprehend | NLP insights from text |
| Amazon Transcribe | Speech to text |
| Amazon Translate | Language translation |
| Amazon Polly | Text to speech |
| Amazon Lex | Conversational chatbots and voice bots |
| Amazon Rekognition | Image and video analysis |
| Amazon Textract | Text, forms, and tables from documents |
| Amazon Personalize | Recommendations |
| Amazon Quick | Agentic analytics and automation workspace for business users |
| Kiro | Agentic, spec-driven IDE for developers |
Commonly confused#
| If the scenario says… | Answer | Not… | Because |
|---|---|---|---|
| Extract fields and tables from scanned invoices | Textract | Comprehend | Comprehend analyzes text you already have; Textract gets it out of documents |
| Detect the sentiment of product reviews | Comprehend | Rekognition | Rekognition is for images and video |
| Turn call recordings into text | Transcribe | Polly | Transcribe is speech → text; Polly is text → speech |
| Read text found in a photo of a street sign | Rekognition | Textract | Rekognition detects text in images and video scenes; Textract is for documents |
| Build a chatbot that books appointments | Lex | Comprehend | Lex manages conversations, intents, and slots |
| One 800 MB video file, result needed in minutes | Asynchronous inference | Batch | One large request, not a whole dataset |
| Score all of last night's transactions | Batch inference | Real-time | Whole dataset, offline |
| Group customers with no predefined categories | Clustering | Classification | Classification needs labels |
| Missing a positive case is dangerous | Recall | Precision | Recall measures missed positives |
| Must explain each loan decision to a regulator | Traditional ML | Foundation model | Explainability and determinism |
Practice questions#
Q1 (matching). Match each requirement to the AWS service that meets it.
| Requirement | |
|---|---|
| 1. Identify the sentiment of social media posts | |
| 2. Extract table data from scanned PDF statements | |
| 3. Convert product descriptions into natural-sounding audio | |
| 4. Moderate user-uploaded images for unsafe content | |
| 5. Produce text transcripts of support calls |
Services: Amazon Polly, Amazon Rekognition, Amazon Comprehend, Amazon Transcribe, Amazon Textract.
Show answer
1 → Comprehend, 2 → Textract, 3 → Polly, 4 → Rekognition, 5 → Transcribe.
- Comprehend analyzes existing text; it can't read a PDF image.
- Textract extracts text, forms, and tables from documents.
- Polly converts text to speech; Transcribe does the reverse.
- Rekognition handles image and video moderation.
Q2 (ordering). Put these ML pipeline stages in order.
- A. Feature engineering
- B. Model monitoring
- C. Data collection
- D. Model training
- E. Evaluation
Show answer
C → A → D → E → B.
Collect data, engineer features from it, train, evaluate on held-out data, and then monitor after deployment. Pre-processing and hyperparameter tuning would fit between these steps if they were listed.
Q3. A media company needs to run an ML model on individual video files of up to 800 MB. Processing each file takes about 20 minutes, and results can be picked up later. Which SageMaker AI inference option fits best?
- A. Real-time inference
- B. Serverless inference
- C. Asynchronous inference
- D. Batch transform
Show answer
Answer: C. Asynchronous inference queues individual requests with large payloads and long processing times, and writes results to S3.
- A and B are built for small payloads and responses within seconds.
- D processes an entire dataset offline; here, files arrive one at a time.
Q4. A hospital is building a model to screen scans for a serious disease. Missing a real case is far worse than a false alarm, which a doctor will review. Which metric should the team prioritize?
- A. Accuracy
- B. Precision
- C. Recall
- D. Mean absolute error
Show answer
Answer: C. Recall measures how many actual positive cases the model catches, so maximizing it minimizes missed cases.
- A is misleading when positives are rare.
- B matters when false positives are costly; here, doctors review false positives.
- D is a regression metric.
Q5. A bank must approve or deny loan applications from structured application data. Regulators require the bank to explain every individual decision. What is the most appropriate approach?
- A. Prompt a large language model on Amazon Bedrock with each application
- B. Train a traditional, interpretable ML model and use SageMaker Clarify for feature attribution
- C. Fine-tune a foundation model on past loan decisions
- D. Use Amazon Comprehend to classify applications
Show answer
Answer: B. Tabular data, a clear target, and a hard explainability requirement point to traditional ML. Clarify explains which features drove each prediction.
- A and C use foundation models, which are harder to explain and less deterministic.
- D analyzes free text, not structured application data.
Q6. A retailer wants to group customers by purchasing behavior. There are no existing customer categories. Which technique fits?
- A. Classification
- B. Regression
- C. Clustering
- D. Reinforcement learning
Show answer
Answer: C. Clustering finds natural groups in unlabeled data.
- A needs predefined labels.
- B predicts a number.
- D learns actions from rewards.
Q7. A logistics company charges shipping fees from a fixed table based on weight and destination zone. A manager suggests using ML to calculate the fees. What should the team do?
- A. Train a regression model on historical fees
- B. Use a rules-based implementation, because the outcome is deterministic
- C. Use a foundation model to read the fee table at runtime
- D. Use clustering to group shipments by zone
Show answer
Answer: B. When a specific, known outcome is required, a rule is cheaper, exact, and explainable. ML predicts; it doesn't guarantee.
- A, C, and D add cost and error to a problem that has an exact answer.
Q8 (multiple response). Which TWO are examples of unsupervised learning?
- A. Detecting anomalous network traffic without labeled examples of attacks
- B. Predicting house prices from past sales
- C. Grouping news articles by topic without predefined topics
- D. Classifying emails as spam using labeled examples
- E. Training a robot arm with rewards
Show answer
Answer: A and C. Both learn structure from unlabeled data.
- B and D are supervised: regression and classification with labels.
- E is reinforcement learning.
Key takeaways#
- AI ⊃ ML ⊃ deep learning ⊃ GenAI. Agentic AI is built on top of a generative model: GenAI creates, agents act.
- Know the four inference types by trigger phrase. Steady low latency → real-time. Intermittent traffic → serverless. One big or slow request → asynchronous. Whole dataset → batch.
- Supervised needs labels; unsupervised doesn't. Foundation models pre-train with self-supervised learning.
- Regression predicts a number, classification a category, clustering groups without labels.
- The managed AI services table is guaranteed matching-question material. Know it cold.
- Regulated plus explainable plus tabular points to traditional ML. Deterministic rules mean no ML at all.
- Pick the metric by the cost of errors. Recall when misses hurt, precision when false alarms hurt, F1 for imbalanced data. Business stakeholders get business metrics.
Next: Domain 2: Fundamentals of GenAI.
Sources#
Originally published at https://iuriio.com/blog/posts/2026/09/aif-c01-part-2-ai-ml-fundamentals
